Deep Transfer Learning for Intelligent Autonomous Vehicles
Autonomous driving has become a very interesting research problem for the deep learning domain. While Intelligent Autonomous Vehicles (IAVs) have developed significantly over the last 10 years, there are still unresolved issues concerning how to transfer knowledge from one driving environment to another. In particular, there is hardly anything known about how to get IAVs trained for driving on one side of the road (e.g., left-hand side in New Zealand and Japan) to right-hand side (e.g., the USA and China). This research describes how a deep learning IAV lane-positioning model can predict the steering angle based on continuous left-hand drive images and velocity inputs for 50 minutes of simulated driving (over 32,000 images) using convolutional neural networks (CNNs). We then examine freezing weights at different layers for successful transfer to right-hand simulated driving (10 minutes and over 7,000 images) and find that the best layers to freeze lie closest to the output layer. By visualizing the effects of weights at different levels, we report that the model shows signs of extracting increasingly relevant features at the higher levels that may help to explain how human drivers transfer knowledge about how to drive on one side of the road to the other. The overall contribution of this thesis is showing how a deep learning IAV model can adhere to lane-positioning by predicting the steering angle and can also transfer knowledge from left hand to right hand drive simulated driving.